SnorkelFinance: leaderboard
Snorkel AI's archived SnorkelFinance benchmark: expert-verified financial questions built from 10-K filings, answered by ReAct agents that must use SQL and code tools without access to the source document. Claude 3.7 Sonnet judges full execution traces, final answers and intermediate steps; score %, higher is better. Frozen roster.
Metric: Score (%, LLM judge over full agent traces). Source: snorkel.ai. Saturation forecast: Estimated already saturated. 28 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5 | 81 | #105 |
| 2 | O3 | 81 | #132 |
| 3 | Gemini 3 Pro | 80.34 | #82 |
| 4 | Claude Opus 4.1 | 80.3 | #146 |
| 5 | GPT-5 Mini | 79.3 | #193 |
| 6 | Claude Opus 4 | 78.3 | #170 |
| 7 | Claude 3.7 Sonnet | 77.9 | #259 |
| 8 | Claude Sonnet 4 | 76.6 | #207 |
| 9 | O4 Mini | 76.6 | #185 |
| 10 | Grok 4 | 74.04 | #189 |
| 11 | Grok 4 Fast (Reasoning) | 73.45 | #266 (Grok 4 Fast) |
| 12 | Kimi K2 (Thinking) | 71.7 | #252 (Kimi K2) |
| 13 | GPT-OSS-120B | 66.6 | #346 |
| 14 | Grok 3 | 65.86 | #313 |
| 15 | O3 Mini | 63.79 | #282 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=snorkelfinance · How It Works · Data refreshed daily, snapshot 2026-11-05.